simple prior use test #461
Conversation
|
Hi, oh great thanks will resolve today! |
|
hi, didnt get through this today, but will look first thing in the morning. |
test/priorusetest.jl
Outdated
|
|
||
| @test isapprox(x0_m, 0.0, atol = 0.1) | ||
| @test isapprox(x1_m, 0.0, atol = 0.1) | ||
| @test isapprox(x2_m, 0.0, atol = 0.1) |
There was a problem hiding this comment.
@Affie, first question: if you are putting two priors at the ends of a 3 variable chain (x0,x1,x2), either at + or - 1 -- why would you expect them to all be at average 0? I would expect (x0,x2) to be at + or -1 one respectively, with x1 floating somewhere near 0 in the middle?
This was based on me reading the code.
There was a problem hiding this comment.
x0 and x2 should be near 0, and only x1 should be "on" 0. Test with loose atol=0.1 bit hacky, but okay :-) x0 and x2 should be around 0.1 or something like that.
There was a problem hiding this comment.
Please note, I changed the standard deviation to 0.01.
So its more like -0.01, 0, 0.01 for x0,x1,x2.
-1/1 <-x0-> 1/0.01 <-x1-> 1/0.01 <-x2-> 1/1
There was a problem hiding this comment.
i'm assuming you did the calculation for multiplying two Gaussians, and found that the new mean should be at +- 0.01? What I mean is if you were to multiply N(-1,0.1) * N(+1,0.1), the answer will be at zero with a new covariance slightly smaller than 0.1.
first test graph
|
|
I don't agree on changes. GraphInit = true
GraphInit = false
Also, forcing it into one clique gives this answer: |
|
adding @tonioteran here for interest sake. I strongly agree that changing the variable ordering should not change the numerical result. Thanks for describing, I definitely need to spend time on this and make sure we understand what is happening here. |
|
I looks like it either has to do with the structure or the priors on the down solve. |
|
acknowledged thanks, definitely will do a deep dive on this soon. |
|
xref #462 |
|
xref #485 |
|
xref #487 |
sub-graphs with only factors on frontals connected
…actor Refactor
fast forward feature branch
Codecov Report
@@ Coverage Diff @@
## master #461 +/- ##
==========================================
+ Coverage 45.04% 46.69% +1.65%
==========================================
Files 29 29
Lines 4234 3735 -499
==========================================
- Hits 1907 1744 -163
+ Misses 2327 1991 -336
Continue to review full report at Codecov.
|
I feel this test should pass